An engineer at a mid-size fintech company described her workflow in a thread earlier this spring: she opens Claude Code in her terminal, gives it the ticket and the relevant codebase context, watches it plan across twelve files and three services, and then presses a key. Cursor takes the plan and executes it in her editor. If there is a discrete function-level task she wants completed in parallel, she routes it to Codex. She did not choose this arrangement from a comparison chart. She arrived at it over several months of trial and then stopped thinking about it.
That pattern has become visible enough, across enough teams, that it now has a name: the composable AI coding stack. SemiAnalysis estimates that Claude Code accounts for roughly 4% of all public GitHub commits as of March 2026, with projections suggesting 20% by year-end. A JetBrains survey of 906 software engineers in February 2026 found it is now the most-loved AI coding tool in actual use at work, at 46%. Meanwhile, Cursor reached $2 billion in annualized revenue by February 2026 — the fastest trajectory from zero to that figure of any B2B software company on record, ahead of Slack, Zoom, and Snowflake at equivalent moments — and is now in talks to raise $2 billion more at a $50 billion valuation, with NVIDIA as a reported strategic participant.
None of these numbers are about the same thing, but they all point at the same structural reality.
What has happened in developer tooling is not consolidation. The conventional story — that the company with the most distribution wins, that GitHub Copilot backed by Microsoft would crowd out the rest — has not played out. Copilot is still the most widely known tool, at 76% awareness in the JetBrains survey. Its growth in actual adoption has stalled. The tools that grew, grew because they solved different problems at different layers of the workflow. Engineers who found them useful composed them rather than choosing between them.
In the first week of April, the market made this architecture explicit. Cursor shipped a rebuilt interface for orchestrating parallel agents. OpenAI published an official Codex plugin designed to run inside Claude Code. Rather than acquiring each other or collapsing into a single surface, the three companies shipped work that made the layering more deliberate. Three tiers settled into place: orchestration at the top — Claude Code, reasoning across the whole codebase and directing agents below it; execution in the editor — Cursor, responding to the orchestrator's plan in real time; and task-running below that — Codex, completing discrete function-level work specified by the layer above. Nobody drew this org chart. Engineers built their way into it.
The reason the stratification happened is worth understanding. The shift from chat to agentic patterns — and what it actually costs to run them — changed the problem that tooling had to solve. A model that suggests the next line of code is an autocomplete engine. A model that can hold an entire codebase in context, plan across a multi-service architecture, write a specification, direct execution agents, and review the output is something closer to a junior technical lead. The problems those two things solve are different, and they belong in different parts of the developer's workflow.
The practical question for a team building software in a different environment — with bandwidth constraints, a smaller tool budget, or in a timezone where US-based support forums are asleep — is which layer actually addresses your specific bottleneck. The orchestration layer pays off when the problem is large-scale codebase reasoning: maintenance of legacy systems, cross-service refactors, architectural planning. The execution layer pays off when the problem is iteration speed on greenfield features. For a team of six, running all three tools simultaneously may be overhead that the work does not justify. The stack is composable; it is not required to be complete.
The engineers who are getting the most out of this tooling are the ones who chose their layer and sized their workflow around it.
The short of it.
Cursor, Claude Code, and Codex have organically settled into a three-layer AI coding stack — orchestration, execution, and task-running — without any acquisition or planned partnership. Cursor hit $2 billion ARR by February 2026 and is in talks for a $50 billion valuation; Claude Code accounts for an estimated 4% of all GitHub commits. The conventional consolidation story did not happen. For builders, the practical question is which of the three layers solves the actual bottleneck in their workflow — not which single tool to adopt, and not whether to run all three at once.